Update IMEX/LLVM versions

Signed-off-by: dchigarev <[email protected]>

Author: dchigarev

cmake/imex-version.txt

@@ -1 +1 @@
-25123cc3692fdbcf837510f39de47ff353d482fc
+0a6d2901990183ecdbab9240dbd8be92036d9c20

cmake/imex.cmake

@@ -14,7 +14,7 @@ if (NOT DEFINED IMEX_INCLUDES)
 
     # TODO: Change to main https://github.com/intel/mlir-extensions when all the
     # required functionality is merged.
-    set(IMEX_URL https://github.com/intel/mlir-extensions)
+    set(IMEX_URL https://github.com/dchigarev/mlir-extensions)
     gc_fetch_content(imex "${IMEX_HASH}" "${IMEX_URL}"
             SET IMEX_CHECK_LLVM_VERSION=ON IMEX_ENABLE_L0_RUNTIME=0
     )

cmake/llvm-version-imex.txt

@@ -1 +1 @@
-3191587666aa3d1e53966bc8876614c7197fac4f
+add6b2f35f2bcf1f59a2ab2d5b3dab124fe0895a

cmake/llvm-version.txt

@@ -1 +1 @@
-f6a756f35a4d0719a96b4e214905369d565d87da
+add6b2f35f2bcf1f59a2ab2d5b3dab124fe0895a

lib/gc/Transforms/GPU/LinalgToXeGPU.cpp

@@ -764,7 +764,9 @@ loadNdDescTiles(PatternRewriter &rewriter, Location loc, ValueRange loadTiles,
   if (vnniConf) {
     vecLoadType = getVnniVector(tileType.getShape(), tileType.getElementType(),
                                 *vnniConf);
-    packedAttr = mlir::UnitAttr::get(rewriter.getContext());
+    if (!transpose_bit) {
+      packedAttr = mlir::UnitAttr::get(rewriter.getContext());
+    }
   }
   SmallVector<Value> loadVec;
   for (auto tile : loadTiles) {
@@ -1165,7 +1167,6 @@ static LogicalResult createDPASKernel(linalg::LinalgOp linalgOp,
   if (vnniFactor == -1)
     return failure();
 
-  VnniConfig vnniConfA{.vnniFactor = vnniFactor, .vnniAxis = 1};
   VnniConfig vnniConfB{.vnniFactor = vnniFactor, .vnniAxis = 0};
 
   // Load A sub-tiles.
@@ -1214,7 +1215,7 @@ static LogicalResult createDPASKernel(linalg::LinalgOp linalgOp,
   // Extract DPAS tiles from loaded sub-tiles.
   TilesArray dpasVecA = extractVecSubTiles(rewriter, loc, loadVecA,
                                            {dimM, kTile}, tileTypeA.getShape(),
-                                           {dpasTileM, dpasTileK}, vnniConfA);
+                                           {dpasTileM, dpasTileK});
   TilesArray dpasVecB = extractVecSubTiles(rewriter, loc, loadVecB,
                                            {kTile, dimN}, tileTypeB.getShape(),
                                            {dpasTileK, dpasTileN}, vnniConfB);
@@ -1629,7 +1630,7 @@ struct LinalgToXeGPU : public gc::impl::LinalgToXeGPUBase<LinalgToXeGPU> {
   using LinalgToXeGPUBase::LinalgToXeGPUBase;
 
   void runOnOperation() override {
-    LinalgToXeGPUOptions options{kTile, stages, dpasTile};
+    LinalgToXeGPUOptions options{kTile, stages, SmallVector<int64_t>{dpasTile.begin(), dpasTile.end()}};
 
     // Run GEMM pattern first to allow fusion with its consumers.
     RewritePatternSet gemmPatterns(&getContext());

lib/gc/Transforms/Pipeline.cpp

@@ -78,10 +78,10 @@ void populateTensorPasses(mlir::OpPassManager &pm) {
 // scf + arith + math + vector + tensor + linalg.brgemm
 void populateVectorPasses(mlir::OpPassManager &pm) {
   // Do promotion for math / arith ops
-  pm.addNestedPass<func::FuncOp>(math::createMathLegalizeToF32());
+  pm.addNestedPass<func::FuncOp>(math::createMathExtendToSupportedTypes());
   // sourceTypeStrs can be extended
   arith::ArithEmulateUnsupportedFloatsOptions options;
-  std::array<std::string, 1> typeStr = {"bf16"};
+  SmallVector<std::string, 1> typeStr = {"bf16"};
   options.sourceTypeStrs = typeStr;
   options.targetTypeStr = "f32";
   pm.addNestedPass<func::FuncOp>(

test/mlir/test/gc/Transforms/GPU/linalg-to-xegpu-dpas-transpose-sep-alloc.mlir

@@ -76,7 +76,7 @@ module {
 // Tile B is already in the correct shape.
 // CHECK:   %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x16xf16> to vector<512xf16>
 // CHECK:   %[[vA_dpas_flat:.+]] = vector.extract_strided_slice{{.*}}: vector<512xf16> to vector<128xf16>
-// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x8x2xf16>
+// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x16xf16>
 // CHECK-COUNT-3: vector.extract_strided_slice
 
 // Perform DPAS computation.

test/mlir/test/gc/Transforms/GPU/linalg-to-xegpu-dpas-transpose-sep.mlir

@@ -75,7 +75,7 @@ module {
 // Tile B is already in the correct shape.
 // CHECK:   %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x16xf16> to vector<512xf16>
 // CHECK:   %[[vA_dpas_flat:.+]] = vector.extract_strided_slice{{.*}}: vector<512xf16> to vector<128xf16>
-// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x8x2xf16>
+// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x16xf16>
 // CHECK-COUNT-3: vector.extract_strided_slice
 
 // Perform DPAS computation.

test/mlir/test/gc/Transforms/GPU/linalg-to-xegpu-dpas-transpose.mlir

@@ -68,7 +68,7 @@ module {
 // Tile B is already in the correct shape.
 // CHECK:   %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x16xf16> to vector<512xf16>
 // CHECK:   %[[vA_dpas_flat:.+]] = vector.extract_strided_slice{{.*}}: vector<512xf16> to vector<128xf16>
-// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x8x2xf16>
+// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x16xf16>
 // CHECK-COUNT-3: vector.extract_strided_slice
 
 // Perform DPAS computation.

test/mlir/test/gc/Transforms/GPU/linalg-to-xegpu-dpas.mlir

@@ -65,7 +65,7 @@ func.func @matmul(%arg0: memref<32x32xf16>, %arg1: memref<32x32xf16>, %arg2: mem
 // Tile B is already in the correct shape.
 // CHECK:   %[[vA_flat:.+]] = vector.shape_cast %[[vA]] : vector<32x16xf16> to vector<512xf16>
 // CHECK:   %[[vA_dpas_flat:.+]] = vector.extract_strided_slice{{.*}}: vector<512xf16> to vector<128xf16>
-// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x8x2xf16>
+// CHECK:   %[[vA_dpas:.+]] = vector.shape_cast %[[vA_dpas_flat]] : vector<128xf16> to vector<8x16xf16>
 // CHECK-COUNT-3: vector.extract_strided_slice
 
 // Perform DPAS computation.